Lessons Learned Using ML for Knowledge Discovery in the Atmospheric Sciences

Although Machine Learning (ML) has demonstrated success on many predictive tasks in the atmospheric sciences (see McGovern et al 2017 for examples on high-impact weather) and more recently at physics-based simulations (refs), using ML to discover new physical knowledge is a different and difficult task. Although quite successful in many other tasks such as game playing, robotics, or image recognition, most ML methods have no built-in understanding of the physics of the real-world. In many cases outside of the atmospheric sciences, this can be overcome through extended simulations. For example, recent work has shown great success in learning new walking gaits for a variety of physical bodies by simply letting the ML experiment inside a simulator for many millions of iterations (ref to recent Deep Mind article). Creating such a simulator for atmospheric sciences is exceptionally difficult (refs) or else it would already be in use with both weather and climate predictions.

In this talk, we discuss many methods that have been used to perform knowledge discovery in the atmospheric sciences, where we work to hypothesize new knowledge given methods that can operate on large data sets. These include the development of novel knowledge discovery techniques as well as novel model interpretation and visualization techniques that are physics aware. We also discuss the challenges of developing such methods especially as a joint effort between ML scientists and physical scientists. In addition, we discuss challenges to adopting ML within atmospheric sciences, particularly as it relates to the interpretability and trust in ML algorithms.